類神經網路對股市反轉點的學習與預測應用

Abstract

類神經網路這個現代資訊科技的神奇功能，以傳統預測方法所無法達到的 非線性調適能力，正在改變整個世界的預測型態，在本研究中，比較類神 經網路與傳統預測方法，証明類神經網路的優越性，並用以解釋本研究採 用類神經網路作為股價的預測方法.再者，本研究提出一個全新的預測股 價反轉點的方法，以這種方式，本研究希望能使投資者從股價指標中獲得 較佳的資訊，並對這些資訊作最充分的利用，同時為了減低在處理股價資 料間的複雜度，研究嘗試分割股價資料為四個組群，再將不同的組群送到 不同的類神經網路，得每個類神經網路能對相同屬性的股價資料，做最佳 的學習.最後，本研究分析上述方法所作的實驗結果，其結果是在預測期 間的平均誤差為2.9天，優於隨機猜測模型的3.64天，但較學習期間的平 均誤差1.3天為差，由此可以顯示出1.本研究的模型是樂觀的。2.本研究 所建立的類神經網路模式太少，應該加多類神經網路模式的數目，以產生 更佳的結果. Neural Network,the magic word of the modern information technology,gets to change the style of our forecasting world now.The ability of nonlinear fitting of Neural Network,of which the traditional forecasting methods are lack ,make it superior to the traditional forecasting methods.To make it clear that how the neural network is superior to the other methods and why we adopt the neural network to the application of stock- price forecasting,in this thesis,we have compared the principle of the traditional forecasting methods and the principle of neural network . Then,we propose a wholly new forecasting model of stock-price to predict when the turn points of stick-price will come.In this way,the investor may get better information from the model than the traditional neural networks' forecasting methods, and they can make the most of the information from the stock-price index.At the same time in order to reduce the complexity of the input data, the thesis tries to divide the input data into four different groups,and make the different groups learned by four different neural networks.Thus , every neural networks makes the most of the information from the data of different groups . This divided- and-conquer method might be the promising method of combining the neural networks and the expert system in the near future. Finally,the thesis analyzes the experiments' result of the preceding method and model.The average error in testing period of the experiment is 2.9 days,which is better than the result from the random predicting model,which average error is 3.64 days,but worse than the average error in learning period,which average error is 1.4 days.This might suggest that (1)the model proposed by this thesis is promising one. (2)the number of the neural networks model base is too little,we might expect a experiment with larger neural networks model base,which will generate much better result.